explain use of scratch diffs in comments

a few layers make use of otherwise unused diffs to accumulate results,
but unless the diffs are cleared in forward this contaminates the
gradients when these layers share a bottom and their backward is
skipped.

src/caffe/layers/accuracy_layer.cu

@@ -71,9 +71,8 @@ void AccuracyLayer<Dtype>::Forward_gpu(
   const int dim = bottom[0]->count() / outer_num_;
   const int num_labels = bottom[0]->shape(label_axis_);
   const int nthreads = outer_num_ * inner_num_;
-  // Since this memory is not used for anything,
-  // we use it here to avoid having to allocate new GPU
-  // memory to accumulate intermediate results in the kernel.
+  // Since this memory is not used for anything, we use it here to avoid having
+  // to allocate new GPU memory to accumulate intermediate results.
   Dtype* acc_data = bottom[0]->mutable_gpu_diff();
   if (top.size() == 1) {
     // simple case - report only global accuracy.
@@ -134,6 +133,7 @@ void AccuracyLayer<Dtype>::Forward_gpu(
       }
     }
   }
+  // Clear scratch memory to prevent interfering with backward (see #6202).
   caffe_gpu_set(bottom[0]->count(), Dtype(0), bottom[0]->mutable_gpu_diff());
 }
 

src/caffe/layers/sigmoid_cross_entropy_loss_layer.cu

@@ -48,9 +48,8 @@ void SigmoidCrossEntropyLossLayer<Dtype>::Forward_gpu(
   // Stable version of loss computation from input data
   const Dtype* input_data = bottom[0]->gpu_data();
   const Dtype* target = bottom[1]->gpu_data();
-  // Since this memory is not used for anything until it is overwritten
-  // on the backward pass, we use it here to avoid having to allocate new GPU
-  // memory to accumulate intermediate results in the kernel.
+  // Since this memory is not used for anything, we use it here to avoid having
+  // to allocate new GPU memory to accumulate intermediate results.
   Dtype* loss_data = bottom[0]->mutable_gpu_diff();
   Dtype* count_data = bottom[1]->mutable_gpu_diff();
   Dtype valid_count;
@@ -70,6 +69,7 @@ void SigmoidCrossEntropyLossLayer<Dtype>::Forward_gpu(
   normalizer_ = get_normalizer(normalization_, valid_count);
   top[0]->mutable_cpu_data()[0] = loss / normalizer_;
 
+  // Clear scratch memory to prevent interfering with backward (see #6202).
   caffe_gpu_set(bottom[0]->count(), Dtype(0), bottom[0]->mutable_gpu_diff());
   caffe_gpu_set(bottom[1]->count(), Dtype(0), bottom[1]->mutable_gpu_diff());
 }

src/caffe/layers/softmax_loss_layer.cu

@@ -36,9 +36,8 @@ void SoftmaxWithLossLayer<Dtype>::Forward_gpu(
   const Dtype* label = bottom[1]->gpu_data();
   const int dim = prob_.count() / outer_num_;
   const int nthreads = outer_num_ * inner_num_;
-  // Since this memory is not used for anything until it is overwritten
-  // on the backward pass, we use it here to avoid having to allocate new GPU
-  // memory to accumulate intermediate results in the kernel.
+  // Since this memory is not used for anything, we use it here to avoid having
+  // to allocate new GPU memory to accumulate intermediate results.
   Dtype* loss_data = bottom[0]->mutable_gpu_diff();
   // Similarly, this memory is never used elsewhere, and thus we can use it
   // to avoid having to allocate additional GPU memory.
@@ -62,6 +61,7 @@ void SoftmaxWithLossLayer<Dtype>::Forward_gpu(
     top[1]->ShareData(prob_);
   }
 
+  // Clear scratch memory to prevent interfering with backward (see #6202).
   caffe_gpu_set(bottom[0]->count(), Dtype(0), bottom[0]->mutable_gpu_diff());
 }